Written Plans Beat Open Prompts
Robinhood's Agentic examples are open-ended. Desks that keep expectancy replace them with closed plans: screen, trigger, size, exit, flat time.
Robinhood's Agentic Trading overview lists useful starting prompts: thematic portfolios, dip-buy rules, target-weight rebalances, bull/bear thesis chat.
Those prompts optimize for conversation. Sitting on your hands optimizes for refusal.
Open prompts invent risk
Ask an agent to "find something good for power hour" and it will. Scanners are green. Stories are cheap. The model has no native cost for a bad fill.
A written plan flips the default:
| Plan step | Job | |-----------|-----| | Screen | Liquid gate before rank-chase | | Trigger | Regime A–D grammar; no trigger, no ticket | | Size | Baseline notional unless checklist is fully green | | Exit | TP and cut beyond the sweep — ladders alone invalid | | Flat clock | Hard time when research continues but place/cancel stops |
The agentic equity scalp playbook is that closed tree, sanitized for public reading.
Plans are stories you can audit
A plan is a story with constraints. After the session you can ask:
- Did we follow the tree or narrate around it?
- Which soft rejects did we talk ourselves out of?
- Did the flat clock fire, or did we "just one more"?
That audit is evidence. It does not require publishing fills or account balances.
Operator override stays visible
If you override the plan with eyes open, log it. The agent should state the conflict before place. Silent override is how itch becomes "the system."
More refusal grammar: When the Agent Should Sit.